The aim of the project is to use a gene regulatory network to aggregate high-level generic behaviors in a modular architecture. In the frame of a study with EADS IW, this architecture is used for missions where a multi-robot system has to protect a base from ennemies incoming from various points of the environment. Therefore, the GRN has to regulate high-level behaviors such as defense, attack, etc. Each behavior computes a direction to perform the action is aimed to (either with a script or with machine learning). The GRN then has to weight these moves to provide the global behavior of the robot. To do so, the GRN has multiple inputs to localize the ennemies, to position itself in the environment and in the swarm. To validate this approach, three experiments were conducted.
This modular architecture regulated by a GRN has first been tested with uncertainty with one defender having to intercept one attacker. In this problem, the position of the attacker is unknown but can be identified time to time thanks to a localization module. The module updates the attacker position in exchange to a high cost. The GRN is here used to regulate the defense and the attack behaviors in addition to regulating the localization requests. It is optimized to minimize the number of requests while keeping a high interception rate. This architecture has been tested with attackers having speeds higher than the defenders' one. The GRN showed more robustness than a script developped by experts, being able to intercept ennemies faster than the defenders. The following videos presents the interception results with one and a team of defenders against one attacker.
In this experience, the GRN has to intercept a sequence of ennemies incoming from various points of the environment. Therefore, the GRN has to regulate four behaviors: defense, attack, regrouping and scattering. After multiple indepandent optimizations by a genetic algorithm, some team strategies emerged. They are presented in the following videos.
In this experience, the architecture has to face a sequence of multiple ennemies incoming simultaneously on the base. To solve this problem, a hierarchization has been designed. A global controller (either a GRN or a script) divides the defenders into teams, each team being allocate to one ennemy. GRNs comming from the previous experience are used to control the robots inside the teams: they create the best formation according to there positions to the ennemy and the size of the team (the GRN is naturally resistant to this kind of changes). A GRN and a script have been designed to divide and allocate the defenders into teams. The videos hereafter present both approaches. The script is currently more robust than GRN in the allocation task but the GRN has a very short training time. However, the results are promizing.
Martin Delecluse (ISAE student)
Sylvain Cussat-Blanc (University of Toulouse - IRIT)
Stéphane Sanchez (University of Toulouse - IRIT)
Nicolas Schneider (EADS Innovation Work)